Join GitHub today
GitHub is home to over 28 million developers working together to host and review code, manage projects, and build software together.Sign up
Add Fit.covar method #1922
This PR adds a
The idea is to develop the fitting framework to have stateless backends, and for
I also made a small change to the
Just now I saw that I forgot to call
Nov 13, 2018
2 of 4 checks passed
I don't recommend to put to much faith on the results returned from Sherpa's covar because the internal implementation relies on very sketchy numerical algorithm. Perhaps, you may want to look into some very good autodiff tools (such as autograd if model func is pure python see: https://github.com/HIPS/autograd for details or adept if in C++ land, see https://github.com/rjhogan/Adept etc...).
I would like to look at Numpy-based autograd or Tensorflow based implementations (e.g. covar estimation and confidence interval computations could probably be added to https://www.tensorflow.org/probability/) which should provide huge advantages by having autograd and in the case of Tensorflow use any compute hardware one has without extra implementation effort. But we don't have the bandwidth to prototype this, at the moment we are just doing Numpy evaluations without gradients.